Documents

market-recon

Try it

GitHub market recon when the user asks for AI money trends, hot niches, competitor signals, or "找风口/赚钱方向调研", especially when sources are restricted to GitHub only or a previous research run was interrupted. Produces an evidence-linked trend report.

What it does

GitHub market recon when the user asks for AI money trends, hot niches, competitor signals, or "找风口/赚钱方向调研", especially when sources are restricted to GitHub only or a previous research run was interrupted. Produces an evidence-linked trend report.

The skill document

Market Recon (GitHub-only)

Run a trend/opportunity recon where every claim traces to a github.com or api.github.com page fetched during this run. Never fill gaps from memory.

Steps

  1. Fix constraints and scope. Default: only github.com and api.github.com may be fetched; no web_search; no writes to user files; report lands in chat. State the constraint set (and any user-added ones) as the first line of the final report. Done when: the constraint set is written down for the report.

  2. Choose staffing. Recon needing ≤8 fetches: run inline. Larger: spawn 3–4 subagents, each task embedding the constraints verbatim plus its own URL list from references/url-sets.md. Do not give subagents file access. Done when: every dispatched task carries the constraint block and its URL list.

  3. Fetch in capped batches. Max 5 parallel fetches (unauthenticated Search API allows ~10 req/min). Set maxChars 12000–20000 per fetch: Search API JSON runs 65–180KB and truncates otherwise. Prefer extractMode: "text" for API JSON, "markdown" for HTML pages. Done when: trending pages and keyword queries returned HTTP 200, or the rate-limit fallback (step 3a) was used.

    3a. Rate-limit fallback. On 403/abuse signals: stop that batch, switch the remaining queries to https://github.com/search?q=...&type=repositories HTML pages (still GitHub), and note the throttling in the report. Done when: throttling is either absent or disclosed.

  4. Salvage interrupted runs before re-running anything. If a subagent run was killed or a turn interrupted: pull each child session's sessions_history (includeTools: true, small limit) and mine toolResult blocks for usable fetch payloads. Re-fetch only gaps the salvage cannot cover. Disclose killed runs and their recovered/lost data in the report. Done when: every usable payload from the dead run is accounted for and nothing was re-fetched needlessly.

  5. Synthesize the report. Group findings by opportunity (风口). Per repo: full name as link, stars, velocity (stars/day or /week), created/pushed dates, one-line monetization angle. Flag repos created <6 months ago with high velocity as "new". Close with 3–5 trend conclusions, each naming its evidence, plus the caveat: GitHub heat proves tool/content demand, not revenue. Done when: every claim has a link fetched this run and killed/missing data is disclosed.

  6. Verify. Re-check the report: zero non-GitHub sources, no unfetched claims, rate-limit and truncation events disclosed, stargazers_count values quoted only when actually seen (never "about"). Done when: the checklist passes; fix the report before sending if not.

Reference

  • references/url-sets.md — proven URL sets (trending pages, zh/en keyword API queries, vertical queries), URL-encoding for CJK queries, and batch/rate-limit notes.

Related skills

GitHub trend intelligence for AI agents — a daily scan of trending new GitHub repos, LLM-analyzed into a structured digest (top categories, language trends,...

GitHub & developer data for AI agents — repo metadata (stars, language, license), repo audits & risk scoring, user profiles, recent releases, and gist conten...

2 installs

Hourly Moltbook community digests, free polling endpoints, plus a pay-per-call post-virality advisor.

46 installs

Use when the user asks for GitHub Trending, GitHub hot repositories, daily/weekly/monthly trending repos, trending projects by programming language, spoken-l...

15 installs